|
AI Builders Digest
Sunday, October 25, 2026
|
|
Security week keeps going, but today it splits into offense and defense at the same time. Aaron Levie wants AI swarms attacking your infrastructure before hackers do. Anthropic is quietly publishing the rulebook for how those swarms should behave. And Replit just hired the team that's been thinking about agent memory longer than almost anyone. These three stories are more connected than they look.
|
|
---
|
|
01
|
Aaron Levie: AI attack swarms are the next big security category
|
|
|
Box CEO Aaron Levie extended his security thesis from yesterday with a specific prediction: agent swarms built for cyber defense are about to be one of the largest categories in AI. His argument is that security is a perfect fit for swarms because the math is simple. More compute means more permutations tested, more code reviewed, more data processed. Every penetration test that never gets scheduled because the team is busy, every patch that sits in a backlog, every misconfigured firewall that nobody had time to audit. Agents have no such calendar.
|
Why it matters: The security teams your company has right now are outnumbered, and the attackers are already running automated systems. If Levie is right, the question for your CISO in 2027 is not "should we use AI agents for security?" but "why aren't we running them yet, and how badly are we losing because of the delay?"
|
|
Source →
|
|
---
|
|
02
|
Replit acquires the Chroma team to solve AI memory
|
|
|
Replit CEO Amjad Masad announced that Jeff and the team from Chroma are joining Replit. Chroma built one of the early tools for RAG, a technique that lets AI systems retrieve and use relevant information rather than relying only on what they learned during training. Masad framed the acquisition around a specific problem: giving AI agents the context and memory they need to do real work, rather than forgetting everything between sessions.
|
Why it matters: Every AI agent that loses the thread when a conversation ends, or can't remember what it did last Tuesday on your codebase, has a memory problem. Chroma's whole existence has been building the plumbing that fixes that. Replit calling this a step toward the "self-driving company" is ambitious language, but the underlying technical gap they're filling is real.
|
|
Source →
|
|
---
|
|
03
|
Codex CLI doubles in usage since DevDay, ships a batch of fixes
|
|
|
Thibault Sottiaux posted a Codex CLI update covering reusable prompt presets, inline model and permission settings, enterprise MCP sign-in, fixes for long session resumptions, and Windows corporate-proxy authentication. The practical news buried in the update: CLI usage has doubled since DevDay.
|
Why it matters: Yesterday's digest noted the teased ship and wondered if the mystery was worth the buildup. The answer appears to be "yes, for developers." 2x growth on a CLI tool is a signal that engineers are running this in real workflows, not just experiments. The Windows corporate-proxy fix in particular means this is reaching into environments where IT departments usually block anything that wasn't approved in 2019.
|
|
Source →
|
|
---
|
|
04
|
Anthropic publishes the safety and privacy details behind Claude Duet
|
|
|
Anthropic's Claude account linked to a detailed look at how Claude Duet handles safety, privacy, and security. The post itself is thin on context in the tweet, but the link points to what appears to be technical documentation covering how the product manages sensitive data.
|
Why it matters: "Claude Duet" is a product worth watching given this week's security thread running through the digest. If Anthropic is publishing detailed safety documentation for it rather than the usual blog-post treatment, that suggests enterprise customers were asking hard questions before signing contracts.
|
|
Source →
|
|
---
|
|
05
|
Kimi K3.1 Pro: a 2.8 trillion parameter open-weights model that beats Fable 5.1
|
|
|
Moonshot AI launched Kimi K3.1 Pro, a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters, with a 1-million-token context window. Latent Space's coverage notes it beats Fable 5.1 on benchmarks, not just Fable 5. Open weights drop November 4. The model's post-training used large-scale reinforcement learning across hundreds of domains, and Moonshot is offering a "Highspeed" API version running 2-3x faster than standard.
|
Why it matters: Two months ago, frontier open-weights models were lagging closed models by roughly six months. That gap is now down to about one month. If you're a developer who chose a closed model because open alternatives couldn't keep up on quality, the calculus for your next project is different than it was in August.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|